Reproduce “Query-efficient scan-and-search watermark localization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Liyuan Wang, Xinsong Ma, Ziye Ma
BELL localizes watermarked passages in mixed-source documents using adaptively chosen token queries. It samples across blocks, directs queries toward promising regions, and applies token-weighted e-BH to e-values, controlling token-weighted false discovery rate under conditional e-validity. In tested settings, at 90% recall on Mistral-7B-Instruct-v0.3 and WikiText-103 with Hu Unbiased, BELL queries 23% of tokens versus WaterSeeker’s 90%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.